CSE Roadmap

Core tier · about 4 hours · 11 parts

Part 4 of 11 · 2 min read

Each dependency is a handoff, and a handoff can lose a fact

A single agent has every earlier fact sitting in its own context; a worker only has whatever fit in the brief it was handed.

Think of a long verbal message passed down a line of people. A single person who did the whole task remembers every detail. But a message relayed through several people can quietly lose a detail at any handoff.

In plain words: redundancy is extra work a worker does because it lacks the whole picture. It is re-discovering things a single agent would already know. The critical path is the longest chain of calls that cannot overlap, so it sets the wall-clock time. A handoff means passing a fact across a dependency edge through a brief. It carries a real chance of losing that fact, and that risk compounds with every edge in the chain.

A worked example. The model assumes each worker takes 25% more steps than a single agent would, purely from lacking the whole picture, and assumes a 3% chance per dependency edge that a handed-off fact is lost. At dependency density 1.0 (18.0 edges on average), that 3%-per-edge loss compounds across the whole chain. That is a meaningful part of why a "robust model" run shows the split's reliability falling to 0.46 against a single agent's 0.70 at that density. The single agent never has to relay anything to itself.

The key lines. From code/mini-fanout/fanout.py:

# from code/mini-fanout/fanout.py
for j in range(n):
    start = p.shared + p.brief + len(deps[j]) * p.result      # worker context: background, its brief, the results it depends on
    c, s, _ = run_agent(p, round(p.steps * (1 + p.redundancy)), p.tools_per_sub, start)
    tokens += c
    ok *= s * (1 - p.handoff_loss) ** len(deps[j])

Read it as: ok is multiplied by (1 - p.handoff_loss) ** len(deps[j]), once per worker. That is raised to the power of how many upstream facts that worker depends on. Reliability does not just add up each worker's own error rate. It also pays a compounding penalty for every fact that had to cross an edge to reach it. That is why dependency-heavy graphs lose more than their step count alone would suggest.

Shared-context tax, redundancy, critical path

The shared-context tax is the background every worker must load into its own context, paid once per worker rather than once. Redundancy is extra work done because a worker lacks the whole picture, such as repeated searches. The critical path is the longest chain of calls that cannot overlap. It sets the wall-clock time.

Example: a worker taking 25% more steps (redundancy) and losing a fact 3% of the time per edge (handoff loss) are two separate costs stacked on the same dependent chain. Together they explain why the model's own "robust model" row shows splitting losing to a single agent specifically on dependent work, not independent work.